SAR image overlap detection method, device, equipment and storage medium

Abstract

The present invention discloses a SAR image overlap detection method, apparatus, device, and storage medium. The method includes: obtaining SAR time-series image data, external DEM data, and external surface classification coverage data of a current detection area; obtaining a regional amplitude map of the current detection area based on the SAR time-series image data, and normalizing the regional amplitude map; obtaining a corresponding interferometric phase map based on the SAR time-series image data, and calculating a coherence coefficient; normalizing the external DEM data, and then band-splicing the normalized regional amplitude map, interferometric phase map, coherence coefficient, normalized external DEM data, and external surface classification coverage data to generate corresponding data blocks, and inputting the data blocks into a preset overlap recognition model to obtain an overlap recognition result for the current detection area. The present invention can improve the accuracy of SAR overlap recognition.

Description

Technical Field

[0001] The present invention relates to the field of synthetic aperture radar interferometry technology, and in particular to a SAR image overlap detection method, device, equipment and storage medium. Background Art

[0002] SAR (Synthetic Aperture Radar) is a radar imaging technology used to obtain various types of information about the Earth's surface. It obtains surface information by transmitting and receiving signals reflected from the ground during satellite orbit. However, due to the relative positioning of objects and the propagation characteristics of radar waves, some objects along the radar propagation path are often partially or completely blocked by other objects, causing the signal in these areas to weaken or disappear, resulting in overlapping or aliasing in the image. "Alignment" has two advantages and disadvantages in SAR image processing. On the one hand, the presence of alignment can seriously interfere with SAR data processing. On the other hand, aligned areas in SAR images can be reconstructed into three dimensions using appropriate de-alignment algorithms to obtain 3D information about the area. This has both advantages and disadvantages. To leverage its advantages and mitigate its disadvantages, accurate identification of aligned areas in SAR images is a prerequisite.

[0003] However, the existing SAR image overlap detection technology has certain scenario limitations in its specific application. It does not make comprehensive use of valuable external auxiliary data, but only uses the original SAR image or homologous data derived from the original SAR image. The recognition system does not add any additional new information, resulting in room for improvement in the overlap recognition accuracy and diversity of application scenarios of the recognition method. Summary of the Invention

[0004] The present invention provides a SAR image overlap detection method, device, equipment and storage medium to solve the technical problem that the existing SAR image overlap detection technology does not make comprehensive use of valuable external auxiliary data and only uses the original SAR image or homologous data derived from the original SAR image, resulting in low overlap recognition accuracy of the recognition method.

[0005] In order to solve the above technical problems, an embodiment of the present invention provides a SAR image overlap detection method, comprising:

[0006] Acquire SAR time-series image data and corresponding external heterogeneous data of the current detection area; wherein the external heterogeneous data includes: external DEM data and external surface classification coverage data;

[0007] Extracting corresponding amplitude information from the SAR time-series image data, obtaining a regional amplitude map of the current detection area based on the extracted amplitude information, and performing normalization processing on the regional amplitude map;

[0008] Performing interference processing on the SAR time series image data to obtain a corresponding interference phase map, and calculating a corresponding coherence coefficient based on the interference phase map;

[0009] Normalizing the external DEM data, then band-joining the normalized regional amplitude map, the interference phase map, the coherence coefficient, the normalized external DEM data, and the external surface classification coverage data to generate corresponding data blocks, and inputting the data blocks into a preset overlap recognition model so that the overlap recognition model recognizes the overlap of the current detection area based on the input data blocks, thereby obtaining an overlap recognition result for the current detection area;

[0010] The overlap recognition model is obtained by training a preset deep learning model with historical SAR time-series image data and corresponding historical external heterogeneous data as input and overlap recognition results of the corresponding area as output.

[0011] As a preferred solution, the generation of the overlay recognition model includes:

[0012] Acquire historical SAR time-series image data and corresponding historical external heterogeneous source data; wherein the historical external heterogeneous source data includes: historical external DEM data and historical external surface classification coverage data;

[0013] Extracting corresponding historical amplitude information from the historical SAR time-series image data, obtaining a historical regional amplitude map of the corresponding area based on the extracted historical amplitude information, and performing normalization processing on the historical regional amplitude map;

[0014] Performing interference processing on the historical SAR time-series image data to obtain a corresponding historical interference phase map, and calculating a corresponding historical coherence coefficient based on the historical interference phase map;

[0015] Normalizing the historical external DEM data, and then band-splicing the normalized historical regional amplitude map, historical interferometric phase map, historical coherence coefficient, normalized historical external DEM data, and historical external surface classification coverage data to generate corresponding data blocks, and using the data blocks as input data samples;

[0016] Marking the overlapping areas in the historical SAR time series image data to generate overlapping labels for the corresponding areas;

[0017] A model training sample is generated according to the input data sample and the overlap label, and then a preset deep learning model is trained according to the model training sample and a preset loss function to generate a corresponding overlap recognition model.

[0018] As a preferred solution, the step of marking the overlapped areas in the historical SAR time series image data and generating overlapped labels for the corresponding areas includes:

[0019] SAR image overlay recognition is performed on the historical SAR time series image data according to different preset overlay recognition methods to obtain different overlay recognition results; wherein the preset overlay recognition methods include: an overlay recognition method based on amplitude and coherence thresholds, a satellite radar line of sight analysis and recognition method, and a DEM slope and aspect analysis method;

[0020] Integrating the overlap recognition results, and then marking the overlap areas in the historical SAR time series image data according to the integrated overlap recognition results to generate overlap labels for the corresponding areas;

[0021] As a preferred solution, generating a model training sample according to the input data sample and the mask label includes:

[0022] Using the input data sample and the masked label as initial model training samples;

[0023] For each sample data in the initial model training sample, randomly select several data enhancement methods from the preset data enhancement methods, and perform data enhancement processing on the sample data in sequence to obtain data-enhanced sample data; wherein the data enhancement methods include: geometric transformation, noise injection, image aliasing, and color transformation;

[0024] All sample data after data augmentation are used as model training samples.

[0025] As a preferred solution, the step of identifying the overlap of the current detection area based on the input data block to obtain the overlap identification result of the current detection area includes:

[0026] Perform image-level data fusion on the input data blocks;

[0027] Perform shallow feature extraction on the fused data, and perform feature fusion on the extracted shallow features;

[0028] According to the fused shallow features, the global features in the fused shallow features are extracted, and then the extracted global features are subjected to feature fusion and feature classification to obtain the overlapping recognition results of the current detection area.

[0029] As a preferred solution, the loss function is:

[0030]

[0031] L branch =α·L CE +β·LIoU ;

[0032] L total =L b1 +L b2 +L b3 ;

[0033] Among them, L CE is the cross entropy loss function, which is used to measure the gap between the overlap probability predicted by the overlap recognition model and the true label; L IoU is the intersection-over-union loss function, which is used to calculate the intersection-over-union ratio of the predicted area of ​​the overlap recognition model and the actual overlap area; N is the number of samples; y i is the true label; p i is the predicted probability of the model; L branch is the deep learning model loss function of the overlap recognition model, which is used to integrate the results of the cross entropy loss function and the intersection-over-union loss function, and calculate the loss function of each branch channel according to the integrated result; α and β are the integrated weights of the cross entropy loss function and the intersection-over-union loss function respectively; L total is the overall loss function of the mask recognition model, which is used to calculate the loss of each model output channel, and then sum the losses of all model output channels, and take the sum as the overall loss of the model, L b1 、L b2 and L b3 are the losses of each model channel respectively.

[0034] As a preferred solution, after obtaining the overlap recognition result of the current detection area, the method further includes:

[0035] According to the overlap recognition result of the current detection area, the precision, recall and intersection-over-union of the overlap recognition result are calculated, and then the model recognition accuracy of the overlap recognition model is evaluated according to the precision, recall and intersection-over-union.

[0036] Based on the above embodiment, another embodiment of the present invention provides a SAR image overlap detection device, comprising: a data acquisition module, an amplitude information extraction module, an interference processing module, and an overlap recognition module;

[0037] The data acquisition module is used to acquire SAR time-series image data of the current detection area and corresponding external heterogeneous data; wherein the external heterogeneous data includes: external DEM data and external surface classification coverage data;

[0038] The amplitude information extraction module is used to extract corresponding amplitude information from the SAR time-series image data, obtain a regional amplitude map of the current detection area based on the extracted amplitude information, and perform normalization processing on the regional amplitude map;

[0039] The interference processing module is used to perform interference processing on the SAR time series image data to obtain a corresponding interference phase map, and calculate a corresponding coherence coefficient based on the interference phase map;

[0040] The overlap recognition module is used to normalize the external DEM data, and then perform band splicing on the normalized regional amplitude map, interference phase map, coherence coefficient, normalized external DEM data and external surface classification coverage data to generate corresponding data blocks, and input the data blocks into a preset overlap recognition model, so that the overlap recognition model can identify the overlap of the current detection area based on the input data blocks and obtain the overlap recognition result of the current detection area; wherein, the overlap recognition model is obtained by training a preset deep learning model with historical SAR time series image data and corresponding historical external heterogeneous data as input and the overlap recognition result of the corresponding area as output.

[0041] Based on the above embodiments, another embodiment of the present invention provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the SAR image overlap detection method described in the above embodiments of the invention is implemented.

[0042] Based on the above embodiment, another embodiment of the present invention provides a storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute the SAR image overlap detection method described in the above embodiment of the invention.

[0043] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0044] The present invention provides a SAR image overlay detection method. Compared with existing overlay identification methods that only use a single original SAR image or homologous data derived from the original SAR image, the present invention not only obtains SAR time-series image data of the current detection area, but also obtains external heterogeneous data corresponding to the current detection area: external DEM data and external surface classification coverage data;

[0045] Then, corresponding amplitude information is extracted from the SAR time-series image data, a regional amplitude map of the current detection area is obtained based on the extracted amplitude information, and the regional amplitude map is normalized; then, interferometric processing is performed on the SAR time-series image data to obtain a corresponding interferometric phase map, and a corresponding coherence coefficient is calculated based on the interferometric phase map; wherein the normalized regional amplitude map, interferometric phase map, and coherence coefficient are homologous data derived from the SAR time-series image data;

[0046] Finally, the external DEM data is normalized, and then the normalized regional amplitude map, interference phase map, coherence coefficient, normalized external DEM data and external surface classification coverage data are band-spliced ​​to generate corresponding data blocks, and the data blocks are input into a preset overlap recognition model, so that the overlap recognition model recognizes the overlap of the current detection area according to the input data blocks, and obtains the overlap recognition result of the current detection area.

[0047] The present invention fuses homologous data with heterogeneous data to construct corresponding data blocks, then inputs the fused data blocks into a preset overlap recognition model, and uses a deep learning model to achieve comprehensive utilization of homologous data and external heterogeneous data, thereby improving the stability of overlap recognition in various application scenarios and the overlap recognition accuracy of SAR. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of a SAR image overlap detection method provided by one embodiment of the present invention;

[0049] Figure 2 This is the overall technical flow chart of the overlapping mask recognition and detection method based on deep learning of multi-source data in the present invention;

[0050] Figure 3 This is a flow chart of making SAR image overlay label samples designed by the present invention;

[0051] Figure 4 This is an example of a sample image of overlapping mask recognition;

[0052] Figure 5 is an example diagram of the masked sample enhancement cell of the present invention;

[0053] Figure 6 The figure is a schematic structural diagram of a SAR image overlap detection device provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0054] To make the objectives, technical solutions, and advantages of this application more clear, the technical solutions in this application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned figure descriptions are intended to cover non-exclusive inclusions.

[0056] In the description of the embodiments of this application, the technical terms "first" and "second" are used only to distinguish different objects and should not be understood to indicate or imply relative importance or implicitly specify the quantity, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, the meaning of "plurality" is more than two, unless otherwise clearly and specifically defined.

[0057] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0058] In the description of the embodiments of this application, the term "and / or" is simply a description of the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent the following three situations: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0059] In the description of the embodiments of the present application, the term "multiple" refers to more than two (including two). Similarly, "multiple groups" refers to more than two groups (including two groups), and "multiple pieces" refers to more than two pieces (including two pieces).

[0060] In the description of the embodiments of the present application, unless otherwise expressly specified or limited, technical terms such as "installed," "connected," "connected," and "fixed" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integration; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; internal connections between two components or interactions between two components. Those skilled in the art can understand the specific meanings of the above terms in the embodiments of the present application based on specific circumstances.

[0061] Example 1

[0062] Please refer to Figure 1 , is a flow chart of a SAR image overlap detection method provided by one embodiment of the present invention, comprising the following specific steps:

[0063] S1. Acquire SAR time-series image data of the current detection area and corresponding external heterogeneous data; wherein the external heterogeneous data includes: external DEM data and external surface classification coverage data;

[0064] Specifically, when performing SAR image overlay detection, the SAR time-series image data and corresponding external heterogeneous data of the current detection area are first acquired. The external heterogeneous data includes external DEM data and external surface classification coverage data. The raw time-series SAR data is then preprocessed, including filtering and denoising, geometric correction, radiometric correction, and image registration.

[0065] S2. extracting corresponding amplitude information from the SAR time-series image data, obtaining a regional amplitude map of the current detection area based on the extracted amplitude information, and performing normalization processing on the regional amplitude map;

[0066] Considering that the overlapped area is the superposition of multiple signals and often has large amplitude values, the amplitude information in the amplitude map can be used as one of the important identification features. Therefore, after basic preprocessing of the regional raw SAR data, the amplitude information is further extracted from the regional raw SAR data to obtain the regional amplitude map, and the amplitude map is normalized as shown in the following formula 1:

[0067]

[0068] Where X' is the normalized data, X max , X minare the maximum and minimum values ​​of the data to be normalized, respectively. The amplitude data to be normalized is processed as in formula (1) above to obtain the normalized value. After normalizing the amplitude information to the maximum and minimum values, it is possible to ensure that the amplitude range is unified to the interval [0, 1] while the original data distribution remains unchanged, which facilitates the subsequent model training and processing.

[0069] S3, performing interference processing on the SAR time series image data to obtain a corresponding interference phase map, and calculating a corresponding coherence coefficient based on the interference phase map;

[0070] Furthermore, considering that overlapping regions exhibit significant differences from non-overlapping regions in SAR image coherence maps and interferometric phase maps, we interferometrically process the master and slave SAR images of the region to generate interferometric phase maps, and calculate the coherence coefficients through coherence calculation. These coherence maps and interferometric phase maps are also used as data for the subsequent deep learning method for overlay identification. These data, including amplitude maps, interferometric phase maps, and coherence maps, are all calculated based on the original SAR images and are therefore considered to be derived from the same source as the original SAR data.

[0071] S4, normalizing the external DEM data, and then band-joining the normalized regional amplitude map, the interference phase map, the coherence coefficient, the normalized external DEM data, and the external surface classification coverage data to generate corresponding data blocks, and inputting the data blocks into a preset overlap recognition model, so that the overlap recognition model recognizes the overlap of the current detection area based on the input data blocks, and obtains the overlap recognition result of the current detection area;

[0072] In addition, it is noted that the formation of overlays in SAR images is often closely related to the terrain undulations of the region. Therefore, external DEM data of the corresponding region is also obtained as auxiliary information for overlay identification. The DEM data is also normalized as described in formula (1) above, unifying the numerical range of the DEM data to [0,1]. However, DEM data lacks information about surface objects, and overlays generated by areas such as trees and buildings cannot be analyzed. Therefore, surface classification coverage data of the region is also obtained to indicate different types of land objects in the region. Tall land objects such as buildings have a greater risk of overlay, while flat land objects basically do not produce overlay. External surface classification coverage data also plays an important role in this. Since the external surface classification coverage data and DEM data do not come from the original SAR image, they can be regarded as external heterogeneous data. This part of data can add additional information entropy to the classification system, thereby helping to improve the accuracy of overlay classification and identification.

[0073] The method proposed in the present invention can comprehensively utilize SAR original data and derived SAR image amplitude maps, coherence maps, phase maps and other homologous data, as well as external heterogeneous DEM data and surface classification coverage data by constructing a multi-branch convolutional overlap recognition deep learning model. In addition, the abstract category information and spatial position information of the extracted overlap features are enhanced by multiple exchanges of features between multiple branches of the model, thereby improving the recognition accuracy and scene robustness of traditional methods for SAR image overlaps, and ensuring the accuracy of overlap recognition while ensuring the accuracy of overlap area position and the integrity of edge information.

[0074] The SAR DataCube is constructed by combining the normalized regional amplitude map, interferometric phase map, coherence coefficient, normalized external DEM data, and external surface classification coverage data. The model structure is then initialized, and the model weights obtained from model training are loaded. The preprocessed SAR DataCube data is then used as the model input, and the model forward propagation process is initiated. This forward propagation process generates a predicted probability map of the corresponding regional overlap. A probability threshold is then set to visualize the probability map and generate the overlap recognition results.

[0075] The overlap recognition model is obtained by training a preset deep learning model with historical SAR time-series image data and corresponding historical external heterogeneous data as input and overlap recognition results of the corresponding area as output.

[0076] Preferably, the generation of the overlay recognition model includes: acquiring historical SAR time-series image data and corresponding historical external heterogeneous data; wherein the historical external heterogeneous data includes: historical external DEM data and historical external surface classification coverage data; extracting corresponding historical amplitude information from the historical SAR time-series image data, obtaining a historical regional amplitude map of the corresponding area according to the extracted historical amplitude information, and normalizing the historical regional amplitude map; interfering the historical SAR time-series image data to obtain a corresponding historical interference phase map, and calculating a corresponding historical coherence coefficient according to the historical interference phase map; Historical external DEM data is normalized, and then the normalized historical regional amplitude map, historical interference phase map, historical coherence coefficient, normalized historical external DEM data and historical external surface classification coverage data are band-spliced ​​to generate corresponding data blocks, and the data blocks are used as input data samples; the overlapping areas in the historical SAR time series image data are marked to generate overlapping labels for the corresponding areas; model training samples are generated according to the input data samples and the overlapping labels, and then a preset deep learning model is trained according to the model training samples and a preset loss function to generate a corresponding overlapping recognition model.

[0077] For details, please refer to Figure 2 , which is the overall technical flow chart of the overlay mask recognition and detection method based on deep learning of multi-source data in the present invention, including the main steps of S1: SAR data preprocessing, S2: construction of overlay mask recognition deep learning sample library, S3: design and training of overlay mask recognition deep learning model, S4: overlay mask recognition based on deep learning model, etc. The above content has explained the overlay mask recognition steps based on deep learning model, that is, the model usage process. The following focuses on the model training process in S1-S3:

[0078] 1. SAR data preprocessing:

[0079] Acquire historical SAR time-series image data and corresponding historical external heterogeneous data; wherein the historical external heterogeneous data include: historical external DEM data and historical external surface classification coverage data; extract corresponding historical amplitude information from the historical SAR time-series image data, obtain a historical regional amplitude map of the corresponding area based on the extracted historical amplitude information, and normalize the historical regional amplitude map; perform interference processing on the historical SAR time-series image data to obtain a corresponding historical interference phase map, and calculate a corresponding historical coherence coefficient based on the historical interference phase map; perform normalization on the historical external DEM data.

[0080] Preferably, the marking of overlapping areas in the historical SAR time series image data and generating overlapping labels for corresponding areas includes: performing SAR image overlapping identification on the historical SAR time series image data according to different preset overlapping identification methods to obtain different overlapping identification results; wherein the preset overlapping identification methods include: an overlapping identification method based on amplitude and coherence threshold, a satellite radar line of sight analysis and identification method, and a DEM slope and aspect analysis method; integrating the overlapping identification results, and then marking the overlapping areas in the historical SAR time series image data according to the integrated overlapping identification results to generate overlapping labels for the corresponding areas;

[0081] Preferably, generating a model training sample based on the input data sample and the overlapping label includes: taking the input data sample and the overlapping label as the initial model training sample; for each sample data in the initial model training sample, randomly selecting several data enhancement methods from the preset data enhancement methods, and performing data enhancement processing on the sample data in turn to obtain data-enhanced sample data; wherein the data enhancement method includes: geometric transformation, noise injection, image aliasing and color transformation; and taking all data-enhanced sample data as model training samples.

[0082] 2. Construction of deep learning sample library for overlapping recognition:

[0083] This method uses a supervised deep learning model framework to detect and identify SAR image overlays. This requires a large number of high-quality overlay recognition samples. A complete overlay recognition sample consists of two parts: input data and corresponding label data. The input data for overlay recognition in this method includes historical SAR raw data, derived homologous SAR amplitude, interferometric phase, and coherence data, and externally acquired heterogeneous data such as regional DEMs and surface classification coverage. The label data corresponds to the overlay region mask in the SAR image data.

[0084] The training of deep learning models requires massive, high-quality, annotated datasets. The method of relying solely on manual visual interpretation for SAR image overlay annotation is inefficient. However, the existing overlay recognition methods have certain limitations in terms of accuracy and stability in diverse SAR application scenarios.

[0085] To address the problems in the production of overlapping identification samples, the present invention adopts a strategy of integrated overlapping identification to produce efficient and accurate overlapping sample labels. The overall concept is to first use multiple overlapping identification methods to mark the overlapping areas in the SAR image separately, and then integrate the overlapping identification results obtained by different methods to determine the final overlapping label. The overlapping label production process integrating multiple methods improves the reliability of overlapping identification labels using a single method, and the automated label recognition production process also improves the efficiency of label production compared to manual visual methods.

[0086] Please refer to Figure 3 , is the flow chart of making SAR image mask label samples designed by the present invention, using Figure 3 The method shown in the figure performs an automated labeling process for overlapped samples. First, the original SAR image is acquired and preprocessed. Then, overlapped detection and recognition of SAR images are performed using a variety of technical methods. The overlapped recognition results obtained under different technical routes are then integrated to obtain the final overlapped label.

[0087] like Figure 3 The overlapping mask identification methods used in the embodiments of the present invention include an amplitude and coherence threshold-based overlapping mask detection method, a satellite line-of-sight analysis method, and a DEM slope and aspect analysis method. It is worth noting that in specific applications, other overlapping mask identification methods can also be flexibly selected based on the scenario.

[0088] (1) Overlap detection method based on amplitude and coherence threshold:

[0089] Since the overlapping areas in SAR images are generated by the superposition of multiple signals, they often have a higher amplitude threshold and a lower coherence threshold, and this characteristic can be used to identify overlapping areas. Figure 2 As shown in

[15] , the amplitude map is first extracted from the original SAR image and the coherence is calculated. Then, the overlap detection method based on the amplitude and coherence thresholds is used to screen points whose amplitude and coherence meet certain conditions by setting the amplitude threshold and the coherence threshold to distinguish the overlap and non-overlap areas in the SAR image.

[0090] (2) Satellite radar line of sight analysis and recognition method: This method determines which parts of the ground objects are blocked and form shadows, and which parts are aliased and form shadows based on the line of sight direction observed by the satellite radar.

[0091] (3) DEM slope and aspect analysis method: Since the slope of the overlapping area is often large, the slope and aspect analysis can be calculated from the DEM data, and a comprehensive analysis of the slope and aspect can be performed to identify the overlapping area.

[0092] Different overlap recognition methods have their own advantages in different application scenarios. In the present invention, the overlap results of SAR images under different methods are obtained through the above-mentioned different overlap recognition methods, and are comprehensively processed through an integration strategy. The integration strategy adopted in the embodiment of the present invention is a voting mechanism. When more than half of the methods identify an area as overlap, the area is marked as an overlap label. However, it is worth noting that when the present invention is implemented, a suitable overlap recognition method can be selected as one of the integration methods according to the specific situation. It is not limited to only using the method in this embodiment, and the integration strategy used can also be adjusted accordingly. For example, according to the overlap recognition effect of different overlap integration sub-methods in different SAR image scenes, the weight of such methods can be adjusted accordingly to achieve the purpose of integrating overlap areas identified by different methods.

[0093] The integrated overlay recognition method designed in the present invention can obtain more accurate and reliable overlay recognition results, and use them as label data for the overlay recognition deep learning training sample library.

[0094] To create the sample library input data, the present invention first extracts amplitude maps from SAR images and calculates coherence maps and interferometric phase information. Simultaneously, DEM data and surface classification coverage data for the corresponding area are obtained externally. Band-splitting is performed to create a SAR-DataCube, which contains diverse overlay information. Compared to using only raw SAR images or data derived from the same source as SAR images as input, the overlay recognition samples of the present invention provide more external heterogeneous data, providing more information for overlay recognition detection.

[0095] Please refer to Figure 4 , is an example diagram of overlapping recognition samples. The final form of an overlapping recognition sample is: ([input data]-[label data]), such as Figure 4 shown.

[0096] The input data is a SAR-DataCube for the same area at different times. Each SAR-DataCube consists of two parts: homologous data, namely amplitude data, phase data, and coherence data obtained after processing the original SAR data, and heterogeneous data, namely regional DEM data and ground classification coverage data obtained externally. The external heterogeneous data adds additional information entropy to the entire system, which is of great significance for overlay recognition. The label data is the overlay area obtained by integrating the above-mentioned multiple overlay recognition methods, and each SAR-DataCube at each moment has a corresponding overlay label. The above is a detailed description of the sample format in the present invention.

[0097] Given the complex and interference-prone imaging conditions of satellite imagery and the diverse types of land features, a rich and diverse collection of overlay recognition samples is crucial for training the overlay recognition model. This can significantly improve the model's overlay recognition capabilities in complex scenarios and its ability to resist noise interference during the recognition process. Real overlay scenes are often complex. Due to differences in satellite imaging perspectives and conditions, overlay regions often experience rotation, geometric deformation, and shadows, all of which interfere with their recognition. Furthermore, overlay regions often blend in with the background, interfering with each other and making them difficult to distinguish. Using only normal overlay samples clearly does not capture the characteristics of overlay in real-world applications.

[0098] The present invention takes into account the relatively fixed and single process of traditional data sample enhancement methods. That is, the data enhancement method and process remain fixed once determined. As a result, even if data enhancement is performed, the generated samples are all obtained through the same fixed enhancement process method, resulting in insufficient sample richness. The present invention specifically designs a data enhancement method using a random sample enhancement pool for the SAR image overlay recognition task. Utilizing this method, the model's accuracy and scene adaptability can be significantly improved in overlay recognition tasks.

[0099] Please refer to Figure 5, is an example diagram of the overlapping sample enhancement pool of the present invention. The present invention constructs a overlapping sample enhancement method pool. According to the main characteristics of the overlapping areas in SAR images, four categories of basic methods are designed in the enhancement method pool, including geometric changes, noise injection, image aliasing, and color enhancement. When data enhancement is applied specifically, one or more methods are first randomly selected from each basic method pool each time, and then the methods selected from each basic method pool are combined in random order, and a sample is enhanced using the combined methods and processes. Through the random enhancement method based on the method pool, it is ensured that the enhancement methods and processes used for each sample in the sample library are different, ensuring the diversity of samples obtained through enhancement. Moreover, in specific use, the sub-methods in each basic method pool can be further adjusted according to different application scenarios to ensure the flexibility of the data enhancement method based on the random method pool.

[0100] The geometric changes in the enhancement method pool of the present invention include rotation, scaling, translation and flipping, etc. Geometric change enhancement can enhance the model's adaptability to changes in the geometric shape of the overlapping area. Random rotation of the image can simulate different satellite imaging angles, so that the model can enhance its recognition ability of overlapping targets in different directions. Random scaling enhancement can change the size of the image to simulate the effects of different observation distances, which helps the model learn multi-scale overlapping targets in SAR images. Translation image enhancement can help the model learn the different position information of the overlapping targets in the image and enhance the model's robustness to changes in target position. Horizontal or vertical flip enhancement can help the model learn the symmetry-related features of the overlapping area. The above geometric enhancement methods can help the model better learn the diverse geometric states of overlapping areas during training.

[0101] Color-related enhancements include adjusting brightness, contrast, and other parameters, as well as performing histogram equalization. SAR images can experience changes in brightness in overlapping areas due to factors such as the imaging environment. By adjusting the brightness, samples can be generated under different environments, helping the model adapt to these changes. In overlapping samples, multiple objects may have similar grayscale levels, making recognition difficult. Increasing contrast can make the boundaries between different objects clearer and improve recognition accuracy. Histogram equalization can evenly distribute the grayscale values ​​of SAR images, improving image clarity. Color enhancement methods are also important for training models to more robustly identify objects under various environmental conditions.

[0102] In addition, the method of injecting noise can enhance the ability of the model to resist noise during the training and learning process. The present invention is based on the noise characteristics of the overlapping area in the real scene, and the sub-noise enhancement methods set include blurring, artifact addition, salt and pepper noise, Gaussian noise and other methods. Artifact addition can simulate the artifacts caused by the imaging process and enhance the adaptability of the model to the overlapping phenomenon under complex backgrounds. The blurring operation can simulate the situation of insufficient image clarity to help the model better handle the overlapping detection phenomenon under blurred image conditions and enhance the overlapping recognition ability in real scenes. Salt and pepper noise simulates the noise interference in the image transmission process by randomly adding salt and pepper noise points to the image. Salt and pepper noise helps to improve the model's overlapping recognition ability under defective images. Gaussian noise enhancement simulates sensor noise by adding random noise that conforms to the Gaussian distribution to the image. Through the above multiple noise injection enhancement methods, the common noise scenes in SAR images can be fully simulated to improve the noise interference resistance of the overlapping recognition algorithm model.

[0103] In addition, the overlapping areas in SAR images have more complex image characteristics. Therefore, image aliasing methods such as Mixup, CutMix, and Mosaic can be used for sample enhancement to effectively improve the generalization ability and robustness of model learning.

[0104] The Mixup method linearly combines two images in a certain ratio to generate new samples. This method can enhance the model's generalization ability and make it more robust to unseen samples. In SAR imagery, overlay samples often contain overlapping information of multiple objects. Using Mixup can simulate the interaction between different objects to a certain extent, helping the model learn richer features and boundaries, thereby improving the SAR overlay recognition ability in complex scenes.

[0105] The CutMi x method randomly crops two images and stitches them together to generate new samples. This method enhances the model's robustness to occlusion or partial occlusion of overlapping regions. In overlapping samples, features may be partially obscured by other objects. CutMi x augmentation effectively simulates this situation, exposing the model to information about missing or overlapping overlapping regions during training, thereby enhancing its adaptability to overlapping samples in various complex scenarios.

[0106] Mosaic technology increases sample diversity by stitching together multiple small image blocks to create new images. SAR image overlay samples often contain varying ground object types, textures, and background noise. Mosaic enhancement technology can generate samples containing ground object information from a variety of overlay areas, enabling the model to learn features from different environments during training, improving recognition performance for diverse overlay samples.

[0107] The above is a brief introduction to the various basic methods in the masking sample enhancement pool of the present invention. In specific applications, they can be modified according to specific circumstances.

[0108] The process of building a deep learning sample library for overlay recognition in this paper involves the integrated creation of overlay recognition labels using multiple methods, preprocessing of input data, and dataset augmentation using a pool of random augmentation methods. This process ultimately builds a high-quality, scenario-rich deep learning sample library for training overlay recognition models.

[0109] Preferably, the overlapping of the current detection area is identified based on the input data block to obtain the overlapping identification result of the current detection area, including: performing image-level data fusion on the input data block data; performing shallow feature extraction on the fused data, and performing feature fusion on the extracted shallow features; extracting global features from the fused shallow features based on the fused shallow features, and then performing feature fusion and feature classification on the extracted global features to obtain the overlapping identification result of the current detection area.

[0110] 3. Design and training of deep learning model structure for overlapping mask recognition:

[0111] This paper designs and proposes a method for identifying overlapping regions in SAR images based on a deep learning model. A convolutional multi-branch deep learning model structure is designed to address the characteristics of overlapping regions in SAR images. This model is characterized by a multiple data fusion process from shallow to deep layers, fusing homologous and heterogeneous information in the SAR DataCube multiple times at different feature levels. High-level abstract features of overlapping regions are extracted through a serial-parallel multi-branch feature extraction module, and finally, the overlapping region is identified and classified using the extracted features using an overlapping feature fusion classification module.

[0112] The model designed in this paper maximizes the use of overlapping features in multi-source input SAR-DataCube information through a multi-level resolution convolution structure and a multi-level feature fusion method, and ensures a good balance between the spatial features of overlapping and the abstract classification features during the processing process, so as to achieve the goal of accurately distinguishing the category information of overlapping targets while ensuring the position accuracy of overlapping target recognition.

[0113] The model accepts SAR-DataCube at multiple moments (time series) after preprocessing as input of different channels. After processing by the model, it also obtains output of multiple channels, where the output of each channel corresponds to the overlap recognition result of the input data at different moments.

[0114] First, the DFM-1 module (Data fusion module-1) performs preliminary fusion of the input data at the image level. The DFM-1 module consists of a fully connected layer (linear layer), an activation function (Leaky ReLu) and a Dropout regularization layer. The module uses the fully connected layer to perform linear changes on the input data of different channels, and in the subsequent model training process, by optimizing the parameters of the fully connected layer, it achieves an adaptive balance between the weights of the input channel data and the importance of the data of different channels. Secondly, the nonlinear expression ability of the model is enhanced by the design of the activation function layer. The present invention adopts the Leaky ReLU activation function. Leaky ReLU can avoid the problem of complete neuron inactivation and improve the model's recognition ability of complex overlapping features. Finally, in order to avoid serious overfitting of the model, the Dropout regularization layer is used for random discarding. The DFM-1 module of the model is composed of the above-mentioned structures. This module plays the role of performing preliminary image-level fusion and nonlinear capability enhancement on the multi-channel overlapping recognition data of the input SAR-DataCube, and uses the Dropout regularization layer to prevent serious overfitting.

[0115] Next, a shallow feature extraction (SFE-1) module (Shallow Feature Extraction-1) extracts shallow features from the output data after the initial fusion processing of the DFM-1 module. Shallow features effectively preserve the spatial positional relationships of overlapping features. This module consists of a convolutional layer, a batch normalization layer, and an activation function layer. The convolutional layer processes the input data through a convolution operation. Convolution considers the neighborhood relationships of pixels within the convolution kernel range of the input data and extracts inter-regional feature dependency patterns of pixels. This avoids the problem of traditional feature extraction methods that can only process single-pixel features and serves as a preliminary function for overlapping feature extraction. The batch normalization layer accelerates model training. By normalizing the input of each layer, batch normalization reduces internal covariate shift in the data and enhances the stability of the network training process. Finally, this module also uses an activation function to enhance the nonlinear representation of the model structure. After the data is processed by the SFE-1 module, shallow overlapping features are obtained. These features effectively preserve the spatial positional relationships of overlapping features, but their abstract classification ability and globality are insufficient.

[0116] The model then uses the FFM-2 (Feature Fusion Module-2) module to further fuse the shallow features obtained by the SFE-1 module. The FFM-2 module first processes the features through global average pooling and max pooling to improve their global expressiveness. It then uses a channel-attention mechanism to fuse the features across channels. Feature extraction is performed using an MLP (Multi-Layer Perceptron) structure consisting of convolutional, activation, and normalization layers, further enhancing the abstract expressiveness of the features.

[0117] The CFE-2 and CFE-3 (Convolutional Feature Extraction) modules then sequentially extract high-level, abstract, and global features related to the overlap. The CFE-2 module is distinguished by its two-branch structure, both consisting of convolution-normalization-activation layers. The difference is that feature data is unsampled upon entering branch 1 of the CFE-2 module, but is downsampled by a factor of 2 upon entering branch 2. This processing ensures that features in branch 1 consistently maintain a high spatial resolution of overlap information, while branch 2 provides stronger global feature representation.

[0118] CFE-3 has a three-branch structure, which is also composed of convolution-normalization-activation function layers. The input data of its three branches is obtained from the output of CFE-2, such as Figure 5 As shown in .

[0119] The input of branch 1 in the CFE-3 module is obtained by upsampling the output data of branch 2 of CFE-2 by a factor of 2 and fusing it with the data of branch 1. This operation enables branch 1 of module 3 to retain a certain spatial resolution information of the overlapped features while also improving the deep abstract expression capabilities.

[0120] The input of CFE-3 branch 2 is obtained by downsampling CFE-2 branch 1 by a factor of 2 and fusing it with the branch 2 data. In this way, CFE-2 branch 1 is used to supplement some of the spatial resolution information lost in branch 2 due to downsampling, and it is used as the input of CFE-3 module branch 2.

[0121] The input of CFE-3 branch 3 is obtained by fusion of CFE-2 branch 1 downsampled by 4 times and branch 2 downsampled by 2 times. This branch has the strongest global feature expression and abstract description capabilities.

[0122] Finally, the features of the three branches output by the CFE-3 module are fused and classified through the FFM-C module 3 (Feature Fusion Module-Classify). The features in branch 1 of the CFE-3 module have the best overlapping spatial resolution features, branch 2 combines spatial resolution features with abstract recognition features, and branch 3 has relatively strong overlapping abstract feature recognition capabilities and global feature expression capabilities. In combination with these characteristics, the FFM-C module design includes a sampling layer. First, features from different branches are sampled to unify the spatial scales between different features. Convolution is then used to further fuse the features. Finally, the softmax activation function is used to map the features to the overlapping classification probability space, resulting in overlapping recognition results for different channels.

[0123] Subsequently, the mask recognition results of different channels are compared with their respective true mask labels to calculate their respective loss values, and the sum of their respective loss values ​​is used as the total loss of the model for optimization.

[0124] The above describes the overall structure of the deep learning model for overlay mask recognition designed in this paper. The model will subsequently be trained based on the constructed overlay mask recognition deep learning sample library. This allows the model to learn the classification features of SAR image overlay mask from the sample library and subsequently identify SAR overlay mask in unknown areas.

[0125] Preferably, the loss function is:

[0126]

[0127] L branch =α·L CE +β·L IoU ;

[0128] L total =L b1 +L b2 +L b3 ;

[0129] Among them, L CE is the cross entropy loss function, which is used to measure the gap between the overlap probability predicted by the overlap recognition model and the true label; L IoU is the intersection-over-union loss function, which is used to calculate the intersection-over-union ratio of the predicted area of ​​the overlap recognition model and the actual overlap area; N is the number of samples; y i is the true label; p i is the predicted probability of the model; L branchis the deep learning model loss function of the overlap recognition model, which is used to integrate the results of the cross entropy loss function and the intersection-over-union loss function, and calculate the loss function of each branch channel according to the integrated result; α and β are the integrated weights of the cross entropy loss function and the intersection-over-union loss function respectively; L total is the overall loss function of the mask recognition model, which is used to calculate the loss of each model output channel, and then sum the losses of all model output channels, and take the sum as the overall loss of the model, L b1 、L b2 and L b3 are the losses of each model channel respectively.

[0130] The loss function is a key component for optimizing performance during model training. It can be used to calculate the gap between the model's predicted overlapped area and the true overlapped area, and to continuously improve model performance by minimizing this gap. This paper sets a combined loss function as the loss during model training based on the characteristics of the overlapped area, enabling the model to obtain accurate overlapped category information while also having accurate spatial location information, as shown in Formula (2):

[0131]

[0132] As shown in the above formula (2-c), the deep learning model loss function designed in the present invention for overlapping recognition is given by the cross entropy loss function L in formula (2-a): CE , and (2-b) intersection-over-union loss function L IoU It consists of two weighted parts, the weights of the two parts are α and β respectively. In addition, N is the number of samples, y i is the true label, p i It is the predicted probability of the model. The cross entropy loss function (2-a) is used to measure the gap between the overlap probability predicted by the model and the true label. By calculating and optimizing the loss function, the model is constrained to obtain a more accurate overlap classification result. Overlap recognition requires not only accurate overlap category information, but also precise spatial position of the overlap. Therefore, the present invention calculates the intersection-over-union ratio of the model prediction area and the true overlap area through the intersection-over-union loss function (2-b) to obtain accurate overlap feature spatial position information while optimizing this part of the loss. Finally, it is integrated by weights α and β in formula (2-c) to calculate the loss function of each branch channel. The loss of each model output channel is calculated by the above loss function, and the sum of the losses of all channels is taken as the overall loss of the model, as shown in formula (2-d).

[0133] The model is trained using mini-batch stochastic gradient descent, updating the network's weight parameters layer by layer through the back-propagation of the gradient. Considering the task characteristics of the mask recognition model, a large learning rate (e.g., 0.01) is set at the initial training stage to enable the model to quickly adapt to the characteristics of the masked samples in the sample library at the beginning of training. Subsequently, during model training, the learning rate is gradually adjusted and reduced based on the validation set loss and accuracy performance to avoid the situation where the model loss continues to oscillate near the optimal point and cannot reach the optimal point. The learning rate adjustment strategy shown above enables the model to achieve faster training convergence speed in the initial training stage and steadily converge to the optimal parameter point as the training process progresses.

[0134] After the model converges and the loss function value stabilizes, save the optimal model weight parameter file at this time to complete the model training.

[0135] Preferably, after obtaining the overlap recognition result of the current detection area, it also includes: calculating the precision, recall rate and intersection-over-union ratio of the overlap recognition result based on the overlap recognition result of the current detection area, and then evaluating the model recognition accuracy of the overlap recognition model based on the precision, recall rate and intersection-over-union ratio.

[0136] 4. Model-based SAR mask recognition:

[0137] After completing the above main steps, a converged deep learning model for overlap recognition is obtained, which has the optimal weight parameters for SAR image overlap recognition. This model can then be used to identify overlapped areas in unknown SAR images. In a specific overlap recognition application, the image to be overlapped is first preprocessed in accordance with the training process. As described in step 1, the SAR image to be predicted is adjusted to a form consistent with the input data during training, and external heterogeneous data is obtained to construct a SAR-DataCube data block. Secondly, the model structure is initialized, and the model weights obtained by training in step 3 are loaded. The preprocessed SAR-DataCube data is used as the model input, and the forward propagation process of the model is started. After the forward propagation of the model, a predicted probability map of the overlapped area is obtained. After that, a probability threshold is set to visualize the probability map to obtain the overlap recognition result.

[0138] Finally, relevant accuracy indicators are selected to evaluate the overlap recognition results of the model recognition. The indicators used in this invention are precision, recall, and intersection over union. Specifically, as shown in formula (3):

[0139]

[0140] In Formula 3, TP is the true positive, representing the number of samples correctly identified as overlapping regions. TN is the true negative, representing the number of samples correctly identified as non-overlapping regions. FP is the false positive, representing the number of samples incorrectly identified as overlapping regions, and FN is the false negative, representing the number of samples incorrectly identified as non-overlapping regions. Using Formula 3, we can calculate precision, recall, and intersection-over-union (IoU) metrics. Precision describes the model's accuracy in identifying overlapping regions, recall indicates the model's recall, or its ability to capture overlapping regions, and iou reflects the model's accuracy in localizing overlapping regions. These metrics provide a comprehensive evaluation of the accuracy and reliability of the model's overlapping recognition.

[0141] It can be seen that the present invention provides a SAR image overlap detection method, which has the following advantages over existing overlap detection methods:

[0142] 1. Existing methods for identifying overlaps either rely solely on single data sources, such as SAR amplitude maps and coherence maps, for overlap identification and analysis, or utilize derived information (coherence coefficient maps or interferometric phases) extracted from raw SAR images for homologous combination for overlap identification, ignoring the role of external heterogeneous data such as DEM data or surface classification coverage. The method for identifying overlaps in SAR images proposed in this paper fuses homologous and heterogeneous data to construct a SAR-Data Cube, and utilizes a deep learning model to achieve comprehensive utilization of external heterogeneous data, thereby improving the stability of overlap identification in various application scenarios.

[0143] 2. In addition, the current overlap recognition and analysis methods are relatively simple. They either use a simple threshold discrimination method or analyze the geometric relationship of the image to identify overlaps. In essence, the features used to distinguish overlaps are relatively simple, and there are some deficiencies in distinguishing overlaps. There is room for further improvement in the accuracy of overlap recognition. The present invention uses a multi-branch convolution operation to capture overlap features in the input data through the designed convolutional deep learning model. Convolution can efficiently utilize the spatial neighborhood correlation between pixels to obtain the overlap features of the image, and through a layer-by-layer deepening structure, gradually obtain overlap features from low-level to high-level, from local to global, and finally use high-level abstract global overlap features to identify overlaps. The overlap recognition method of the present invention makes full use of the deep abstract overlap features, which improves the detection accuracy of overlaps compared to traditional methods.

[0144] 3. In order to ensure that the constructed deep learning model is fully and reliably trained and learns about the diverse features of overlap in various complex scenarios, the present invention first designs a multi-method integrated overlap label production process. This method ensures that the acquired overlap labels have high reliability. Secondly, a overlap sample enhancement method based on a random enhancement method pool is designed. This enhancement method ensures that the data enhancement methods used for each overlap sample are not exactly the same. Compared with the single and fixed data enhancement method of the traditional process, it guarantees the diversity of enhanced samples and the diversity of enhancement methods to the greatest extent. The above method can construct a rich and stable overlap recognition sample library to improve the training effect of the overlap recognition model.

[0145] Example 2

[0146] Please refer to Figure 6 , is a schematic structural diagram of a SAR image overlap detection device provided by one embodiment of the present invention, the device comprising: a data acquisition module, an amplitude information extraction module, an interference processing module, and an overlap recognition module;

[0147] The data acquisition module is used to acquire SAR time-series image data of the current detection area and corresponding external heterogeneous data; wherein the external heterogeneous data includes: external DEM data and external surface classification coverage data;

[0148] The amplitude information extraction module is used to extract corresponding amplitude information from the SAR time-series image data, obtain a regional amplitude map of the current detection area based on the extracted amplitude information, and perform normalization processing on the regional amplitude map;

[0149] The interference processing module is used to perform interference processing on the SAR time series image data to obtain a corresponding interference phase map, and calculate a corresponding coherence coefficient based on the interference phase map;

[0150] The overlap recognition module is used to normalize the external DEM data, and then perform band splicing on the normalized regional amplitude map, interference phase map, coherence coefficient, normalized external DEM data and external surface classification coverage data to generate corresponding data blocks, and input the data blocks into a preset overlap recognition model, so that the overlap recognition model can identify the overlap of the current detection area based on the input data blocks and obtain the overlap recognition result of the current detection area; wherein, the overlap recognition model is obtained by training a preset deep learning model with historical SAR time series image data and corresponding historical external heterogeneous data as input and the overlap recognition result of the corresponding area as output.

[0151] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement the present invention without inventive effort.

[0152] Those skilled in the art will clearly understand that for the sake of convenience and brevity, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0153] Example 3

[0154] Accordingly, an embodiment of the present invention provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the SAR image overlap detection method described in the above-mentioned embodiment of the invention is implemented.

[0155] The electronic device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The device may include, but is not limited to, a processor and a memory.

[0156] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the device and connects various parts of the entire device using various interfaces and lines.

[0157] Example 4

[0158] Accordingly, an embodiment of the present invention provides a storage medium, which includes a stored computer program. When the computer program is executed, the device where the storage medium is located is controlled to execute the SAR image overlap detection method described in the above embodiment of the invention.

[0159] The memory can be used to store the computer program, and the processor realizes various functions of the device by running or executing the computer program stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0160] The storage medium is a computer-readable storage medium, and the computer program is stored in the computer-readable storage medium. When the computer program is executed by the processor, it can implement the steps of the above-mentioned various method embodiments. The computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0161] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A SAR image overlap detection method, characterized in that: include: Acquire SAR time-series image data and corresponding external heterogeneous data of the current detection area; wherein the external heterogeneous data includes: external DEM data and external surface classification coverage data; Extracting corresponding amplitude information from the SAR time-series image data, obtaining a regional amplitude map of the current detection area based on the extracted amplitude information, and performing normalization processing on the regional amplitude map; Performing interference processing on the SAR time series image data to obtain a corresponding interference phase map, and calculating a corresponding coherence coefficient based on the interference phase map; Normalizing the external DEM data, then band-joining the normalized regional amplitude map, interference phase map, coherence coefficient, normalized external DEM data, and external surface classification coverage data to generate corresponding data blocks, and inputting the data blocks into a preset overlap recognition model so that the overlap recognition model performs image-level data fusion on the input data blocks; Perform shallow feature extraction on the fused data and perform feature fusion on the extracted shallow features; when performing shallow feature extraction, the pixel domain relationship within the convolution kernel range of the input data is considered, and the feature dependency pattern between pixel regions is extracted; Based on the fused shallow features, the global features in the fused shallow features are extracted through the CFE-2 module and the CFE-3 module in turn. The extracted global features are then fused and classified to obtain the overlapped recognition results of the current detection area. The CFE-2 module consists of two branches. Feature data is not sampled when entering CFE-2 branch 1, and is downsampled by a factor of 2 when entering CFE-2 branch 2. The CFE-3 module consists of three branches. The input of CFE-3 branch 1 is obtained by upsampling the output data of CFE-2 branch 2 by a factor of 2 and fusing it with the output data of CFE-2 branch 1. The input of CFE-3 branch 2 is obtained by downsampling the output data of CFE-2 branch 1 by a factor of 2 and fusing it with the output data of CFE-2 branch 2. The input of CFE-3 branch 3 is obtained by downsampling the output data of CFE-2 branch 1 by a factor of 4 and fusing it with the output data of CFE-2 branch 2 by a factor of 2. The overlap recognition model receives data blocks at multiple moments as inputs of different channels, and obtains outputs of multiple channels as the overlap recognition results after processing; The overlap recognition model is obtained by training a preset deep learning model with historical SAR time-series image data and corresponding historical external heterogeneous data as input and overlap recognition results of the corresponding area as output.

2. The SAR image overlap detection method according to claim 1, wherein: The generation of the overlap recognition model includes: Acquire historical SAR time-series image data and corresponding historical external heterogeneous source data; wherein the historical external heterogeneous source data includes: historical external DEM data and historical external surface classification coverage data; Extracting corresponding historical amplitude information from the historical SAR time-series image data, obtaining a historical regional amplitude map of the corresponding area based on the extracted historical amplitude information, and performing normalization processing on the historical regional amplitude map; Performing interference processing on the historical SAR time-series image data to obtain a corresponding historical interference phase map, and calculating a corresponding historical coherence coefficient based on the historical interference phase map; Normalizing the historical external DEM data, and then band-splicing the normalized historical regional amplitude map, historical interferometric phase map, historical coherence coefficient, normalized historical external DEM data, and historical external surface classification coverage data to generate corresponding data blocks, and using the data blocks as input data samples; Marking the overlapping areas in the historical SAR time series image data to generate overlapping labels for the corresponding areas; A model training sample is generated according to the input data sample and the overlap label, and then a preset deep learning model is trained according to the model training sample and a preset loss function to generate a corresponding overlap recognition model.

3. The SAR image overlap detection method according to claim 2, wherein: The step of marking the overlapped areas in the historical SAR time series image data and generating overlapped labels for the corresponding areas includes: SAR image overlay recognition is performed on the historical SAR time series image data according to different preset overlay recognition methods to obtain different overlay recognition results; wherein the preset overlay recognition methods include: an overlay recognition method based on amplitude and coherence thresholds, a satellite radar line of sight analysis and recognition method, and a DEM slope and aspect analysis method; The overlap recognition results are integrated, and then the overlap regions in the historical SAR time series image data are marked according to the integrated overlap recognition results to generate overlap labels for the corresponding regions.

4. The SAR image overlap detection method according to claim 2, wherein: Generating a model training sample according to the input data sample and the mask label includes: Using the input data sample and the masked label as initial model training samples; For each sample data in the initial model training sample, randomly select several data enhancement methods from the preset data enhancement methods, and perform data enhancement processing on the sample data in sequence to obtain data-enhanced sample data; wherein the data enhancement methods include: geometric transformation, noise injection, image aliasing, and color transformation; All sample data after data augmentation are used as model training samples.

5. The SAR image overlap detection method according to claim 2, wherein: The loss function is: ； ； ； ； in, is a cross entropy loss function, which is used to measure the gap between the overlap probability predicted by the overlap recognition model and the true label; is the intersection-over-union loss function, which is used to calculate the intersection-over-union ratio of the predicted area of ​​the overlap recognition model and the actual overlap area; N is the number of samples; is the true label; is the predicted probability of the model; is a deep learning model loss function of the overlapped recognition model, used to integrate the results of the cross entropy loss function and the intersection-over-union loss function, and calculate the loss function of each branch channel according to the integrated result; and are the comprehensive weights of the cross entropy loss function and the intersection-over-union loss function respectively; is the overall loss function of the mask recognition model, which is used to calculate the loss of each model output channel, and then sum the losses of all model output channels, and use the sum as the overall loss of the model. 、 and are the losses of each model channel respectively.

6. The SAR image overlap detection method according to claim 1, wherein: After obtaining the overlap mask recognition result of the current detection area, the following steps are also included: According to the overlap recognition result of the current detection area, the precision, recall and intersection-over-union of the overlap recognition result are calculated, and then the model recognition accuracy of the overlap recognition model is evaluated according to the precision, recall and intersection-over-union.

7. A SAR image overlap detection device, characterized in that: The SAR image overlap detection method according to any one of claims 1 to 6 comprises: a data acquisition module, an amplitude information extraction module, an interference processing module, and an overlap recognition module; The data acquisition module is used to acquire SAR time-series image data of the current detection area and corresponding external heterogeneous data; wherein the external heterogeneous data includes: external DEM data and external surface classification coverage data; The amplitude information extraction module is used to extract corresponding amplitude information from the SAR time-series image data, obtain a regional amplitude map of the current detection area based on the extracted amplitude information, and perform normalization processing on the regional amplitude map; The interference processing module is used to perform interference processing on the SAR time series image data to obtain a corresponding interference phase map, and calculate a corresponding coherence coefficient based on the interference phase map; The overlap recognition module is used to normalize the external DEM data, and then perform band splicing on the normalized regional amplitude map, interference phase map, coherence coefficient, normalized external DEM data and external surface classification coverage data to generate corresponding data blocks, and input the data blocks into a preset overlap recognition model, so that the overlap recognition model recognizes the overlap of the current detection area based on the input data blocks, and obtains the overlap recognition result of the current detection area; The overlap recognition model is obtained by training a preset deep learning model with historical SAR time-series image data and corresponding historical external heterogeneous data as input and overlap recognition results of the corresponding area as output.

8. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for detecting SAR image overlap according to any one of claims 1 to 6 is implemented.

9. A storage medium, characterized in that: The storage medium includes a stored computer program, wherein when the computer program is executed, the device where the storage medium is located is controlled to execute the SAR image overlap detection method according to any one of claims 1 to 6.

Images